Karin Pelka
Research Interests
Make your next move.
Prepare your application
Review drafts before you send them.
Explore related searches
Articles (3)
Year:
2025
Tissues consist of multi-cellular neighborhoods in which different cell types express correlated gene programs due to shared signaling environments. Methods for identifying these spatial neighborhoods may be powerful, but currently do not scale to existing data sets of millions of cells and often artificially divide tissues into distinct neighborhoods with hard borders. To better identify multi-cellular microenvironments with shared gene programs in large-scale spatial genomics data, we developed a method that combines nonnegative matrix factorization (NMF) with Gaussian smoothing across cells in space. Our spatially-aware dimension reduction, neighborhood NMF (NNMF), identifies known and unknown interactions among diverse cell types organized into complex patterns, from localized structures to broad tissue regions. NNMF has many advantages over currently available methods, including the ability to run on modern large-scale data with thousands of features and multiple tissue samples and with millions of cells. Furthermore, our method is based on probabilistic NMF, which produces soft clusters of landscape signatures that can be understood as overlapping spatially-organized multicellular gene expression programs, allowing more biologically-complete interpretations than overly-simplistic hard clustering methods. In a benchmark dataset of a diverse set of spatial gene expression data with expert tissue labels, compared against related methods, NNMF with K-nearest neighbors clustering shows excellent performance even on hard clustering tasks. On MERFISH human colorectal cancer data, NNMF identifies several immunologically relevant multicellular interaction networks and scales to these data sets with million of cells. NNMF is implemented as an R package available at https://github.com/ragnhildlaursen/NNMF .
Year:
2025
While often represented as static entities, gene networks are highly context-dependent. Here, we developed a multi-task learning strategy to yield context-specific representations of gene network dynamics. We assembled a corpus comprising ~103 million human single-cell transcriptomes from a broad range of tissues and diseases and performed a two stage pretraining, first with non-malignant cells to generate a foundational model and then with continual learning on cancer cells to tune the model to the cancer domain. We performed multi-task learning with the foundational model to learn context-specific representations of a broad range of cell types, tissues, developmental stages, and diseases. We then leveraged the cancer-tuned model to jointly learn cell states and predict tumor-restricting factors within the colorectal tumor microenvironment. Model quantization allowed resource-efficient fine-tuning and inference while preserving biological knowledge. Overall, multi-task learning enables context-specific disease modeling that can yield contextual predictions of candidate therapeutic targets for human disease.
Year:
2024
Collaborators (2)
Barbara Engelhardt
Professor
Stanford University
Ravi Madduri
Senior Scientist
Argonne National Laboratory
Make your next move.
Prepare your application
Review drafts before you send them.